Texture recognition based on magnetostrictive tactile sensor array and convolutional neural network

Abstract

Tactile sensors are key devices in surface information perception for robots that can recognize the surface texture of fabrics with different materials and winding patterns in unstructured environments, thus helping the robot process fabrics more effectively. In this study, a magnetostrictive tactile sensor array was designed and loaded onto the robotic fingertips, and the output voltage waveform was obtained by manipulating the sensor array to slide in contact with fabrics. The output voltage waveform diagram was normalized to build the FTS-15 tactile texture dataset. The convolutional neural network ResNet-18 model was built to pre-process the dataset, and the accuracy of recognizing 15 fabrics reached 97.95%. The results show that this texture recognition method can be effectively applied to the field of fabric texture recognition.